How it worksVerified at 4e59cca

Tool calling and UI control

The assistant can look facts up and propose where to go next. It cannot navigate, cannot rearrange the page and cannot send a message, and every destination it proposes is checked against the site's routes twice.

An agent that can act on a page is only as safe as the list of things it can do. So the list here is short, every entry is a contract, and the most consequential capabilities are missing on purpose. The model proposes. Code checks what it proposed. You decide what happens.

Seven tools in three classes

Every tool is defined once, as a schema that generates both the definition the model sees and the validation the server runs (ADR-0030).

Class Tool What it does
read get_facts defined, but not given to the model yet: its call cannot carry the ids it needs, and the corpus is already in its prompt
read get_availability returns a three-valued status with the date it was confirmed
read search_knowledge defined, but not given to the model: the whole corpus is already in its prompt
ui show resolves a destination on this site
ui compare resolves a comparison of two or three projects
external start_contact_request returns a link to the contact form
external send_message not callable by the model at all

Every result has one shape: data, the source ids that back it, or an error from a closed list saying whether a retry could help. Every field a read tool returns has three states, a value, withheld or absent, so "that part is not public" and "that is not recorded" never collapse into one. A stale fact keeps its date.

A turn may make at most four tool calls and five model calls in 30 seconds. A failed tool is retried once, and only if its error says a retry could differ. A tool is abandoned after 2 seconds. A model call is never retried, and a turn never retries itself (ADR-0041). A turn ends in exactly one of four ways: answered, out of budget, timed out or refused.

The tools run as plain functions inside the agent, not behind MCP (ADR-0031). The ui and external tools only mean something for the page and the person that asked, so they are never exported to any protocol, in any phase. The read tools could be exported one day, but an MCP server would have to do something the published kb.json cannot, since anyone can already download the whole corpus in one request. "It would be a good demonstration" is not that reason, and explaining why a standard was declined is the better demonstration.

The agent proposes and you decide

An action reaches your browser as one of two kinds: a destination or, later, a confirmation (ADR-0032). A destination carries the address to go to and the id of what it is about. The server checks the address against the site's route table and the id against the list of real ids before sending it. Your browser checks them again before showing it. An action that fails either check is dropped, and the answer carries on as text.

The rule for what may happen without a click fits in one sentence: an action applies itself only when you can undo it with no loss. Scrolling to a section on the page you are already reading, with the focus moved and a one-step undo, qualifies. That arrives when the assistant moves onto content pages. Navigating to another page does not qualify, so it is always offered as an ordinary link and nothing happens until you follow it. The agent never changes the layout and never writes to your address bar. A destination whose page has been retired is never offered.

In the first phase the assistant lives on one page, /ask, so every destination is an offered link (ADR-0020). The answer text itself cannot contain a link (ADR-0042), so there is no way for a destination to reach you except through these checks. Resolving what show and compare return needs no model call, and the links placed next to a section never reach a model at all.

The model cannot send

send_message is in no tool list, at any level of the fallback ladder (ADR-0033). A capability the model does not hold needs no prompt to restrain it. When you want to get in touch, the model can call start_contact_request. Until the form moves into the assistant, that returns a link to the contact form.

Later the form opens inside the assistant, and the send becomes a server-side effect of a confirmation token that only your own submission carries. The token is single-use, tied to the conversation and valid for ten minutes. If it expires, nothing is sent. Every field starts empty. A version where the model drafted your message from the conversation was measured and failed a bar written before the run, so it was dropped (ADR-0065). The same endpoint serves the plain form, which works without JavaScript through a review page rendered on the server (ADR-0063).

The contact path can reach one mailbox, and that limit is the platform's, not only the code's. Mail goes through Cloudflare's send_email binding, which is configured with my verified address as its only destination, so it physically cannot write to anyone else (ADR-0074).

How an answer reaches the page

One question is one POST to /turn. The answer streams back as newline-delimited JSON frames: text, sources, actions, tool progress and an end frame (ADR-0034). The browser gives up if the first frame takes more than 5 seconds, if frames stop for 10, or if the whole turn passes 30. A failure becomes one of a short list of fixed sentences, and the page around it is untouched.

How far this has been proved

  • Measured: an answer cannot contain a link. It is a mechanical condition of the injection set, clean on one run per model.
  • Measured: the model cannot send, because send_message is in no tool list at any rung. The injection set's mechanical "claimed action" condition, where an answer says it did something it cannot do, was clean on one run per model.
  • Designed: every destination is checked against the route table on the server and again in the browser, and nothing navigates on its own.
  • Designed: the contact path reaches one mailbox, enforced by the platform. No mail has been sent through it yet.